[Bug 266013] Reported system memory decreases on large systems
- Reply: bugzilla-noreply_a_freebsd.org: "[Bug 266013] Reported system memory decreases on large systems"
- Reply: bugzilla-noreply_a_freebsd.org: "[Bug 266013] Reported system memory decreases on large systems"
- Reply: bugzilla-noreply_a_freebsd.org: "[Bug 266013] Reported system memory decreases on large systems"
- Reply: bugzilla-noreply_a_freebsd.org: "[Bug 266013] Reported system memory decreases on large systems"
- Go to: [ bottom of page ] [ top of archives ] [ this month ]
Date: Wed, 24 Aug 2022 08:37:14 UTC
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=266013 Bug ID: 266013 Summary: Reported system memory decreases on large systems Product: Base System Version: CURRENT Hardware: Any OS: Any Status: New Severity: Affects Only Me Priority: --- Component: kern Assignee: bugs@FreeBSD.org Reporter: thj@FreeBSD.org Created attachment 236086 --> https://bugs.freebsd.org/bugzilla/attachment.cgi?id=236086&action=edit Plot of memory reported by system On large database systems we have, the memory reported by top (as reported by the vm sysctls) decreases over time. This leads to the appearance that FreeBSD is loosing pages. We are experiencing this on a number of systems running FreeBSD 13, with memories in the Terrabyte ranges (systems have 4TB, 3.2TB), and multiple NUMA domains (2 and 4). Over time the missing pages increase until they account for ~50% of installed system memory with roughly 1.5TB missing on the 3.2TB machines. Over time the memory fields report by top (active, inactive, laundry, wired, buf, free) stopping summing up to the amount of memory in the machine. Included is a plot showing the total memory that would be reported by top as a percentage of the total pages the system has. We have captured the output of `sysctl vm` on effected systems over long periods, an hour with 1 second sampling (upper plot) and 24 hours with 1 minute sampling (lower two plots). While debugging this effect we patched the vm system to report the per domain page_count. The upper plot in the figure shows this information for each of the memory domains in that system as well. These systems are part of a Galera MySQL cluster. When the Galera processes are restarted the available memory in the system returns. -- You are receiving this mail because: You are the assignee for the bug.